Note

Transcription

Note
Big ideas
•
•
•
•
Straightforward discriminative methods are powerful
Spot locally discriminative properties
•
•
•
skin
faces
motion
Pool these for detection
Reason about relations
Finding skin
•
Skin has a very small range of (intensity independent)
colours, and little texture
•
•
Compute an intensity-independent colour measure, check if colour is in
this range, check if there is little texture (median filter)
See this as a classifier - we can set up the tests by hand, or learn them.
Histogram based classifiers
•
•
•
Represent class-conditional densities with histogram
Advantage:
•
estimates become quite good
(with enough data!)
•
Disadvantage:
•
Histogram becomes big with high dimension
but maybe we can assume feature independence?
•
Histogram classifier for skin
Figure from Jones+Rehg, 2002
Figure from Jones+Rehg, 2002
Curse of dimension - I
•
This won’t work for many features
•
•
try R, G, B, and some texture features
too many histogram buckets
Finding faces
•
•
•
Faces “look like” templates (at least when they’re frontal).
General strategy:
•
•
•
search image windows at a range of scales
Correct for illumination
Present corrected window to classifier
•
•
•
•
How corrected?
What features?
What classifier?
what about lateral views?
Issues
The Thatcher Illusion
Figures by Henry Rowley,
http://www.cs.cmu.edu/~har/puzzle.html
The Thatcher Illusion
Figures by Henry Rowley,
http://www.cs.cmu.edu/~har/puzzle.html
Naive Bayes
•
Previously, we detected with a likelihood ratio test
P (features|event)
> threshold
P (features|not event)
•
Now assume that features are conditionally independent
given event
P (f0 , f1 , f2 , . . . , fn |event) = P (f0 |event)P (f1 |event)P (f2 |event) . . . P (fn |event)
Naive Bayes
•
•
(not necessarily perjorative)
Histogram doesn’t work when there are too many features
•
•
•
•
•
the curse of dimension, first version
assume they’re independent conditioned on the class, cross fingers
reduction in degrees of freedom
very effective for face finders
relations may not be all that important
very effective for high dimensional problems
bias vs. variance
•
•
Work by Schneiderman and Kanade,
http://www.cs.cmu.edu/afs/cs.cmu.edu/user/hws/www/hws.html
Work by Schneiderman and Kanade,
http://www.cs.cmu.edu/afs/cs.cmu.edu/user/hws/www/hws.html
Many more face finders on the face detection home page
http://home.t-online.de/home/Robert.Frischholz/face.htm
From “Gender Classication with
Support Vector Machines”
Baback Moghaddam
Ming-Hsuan Yang, MERL TR
Neural networks
•
Logistic regression heavily generalized
Training
•
Choose parameters to minimize error on training set
•
•
Stochastic gradient descent, computing gradient using
trick (backpropagation, aka the chain rule)
Stop when error is low, and hasn’t changed much
Rowley-Baluja-Kanade face finder (1)
Figure from “Rotation invariant neural-network based face detection,”
H.A. Rowley, S. Baluja and T. Kanade, Proc. Computer Vision and Pattern Recognition,
1998, c 1998, IEEE as shown in Forsyth and Ponce, p589
Figure from “Rotation invariant neural-network based face detection,”
H.A. Rowley, S. Baluja and T. Kanade, Proc. Computer Vision and Pattern Recognition,
1998, c 1998, IEEE as shown in Forsyth and Ponce, p589
Pedestrian detection with an SVM
Dalal+Triggs 05
Features
Dalal+Triggs 05
Performance
Dalal+Triggs 05
Face Recognition
•
•
•
Whose face is this? (perhaps in a mugshot)
Issue:
•
•
What differences are important and what not?
Reduce the dimension of the images, while maintaining the “important”
differences.
One strategy:
•
•
Principal components analysis, then nearest neighbours
Many face recognition strategies at http://www.cs.rug.nłusers/peterkr/
FACE/face.html
Curse of dimension-II
•
•
•
General phenomenon of high dimensions
•
volume is concentrated at the boundary
•
even Gaussians
where probability is concentrated further and further from the mean
and covariance has too many parameters
dodge: assume covariance is diagonal
Parameter estimation is hard for high dimensional
distributions
•
•
•
Idea: reduce the dimension of the feature set
•
•
Principal components
Linear discriminants
Principal components for face images, from
http://vismod.www.media.mit.edu/vismod/demos/facerec/basic.html
Linear discriminant analysis
•
•
Principal components do not preserve discrimination
•
so we could have features that don’t distinguish, see picture
•
Choose linear features so that
between class variation is big compared to within class variation
between class variation
covariance of class means
within class variation
class covariance
Assume (pretend) class conditional densities are normal,
with the same covariance
•
•
•
•
•
First two canonical variates for well
known image collection
More complex template matching
•
•
Encode an object as a set of patches
•
•
centered on interest points
match by
voting
spatially censored voting
inference on a spatial model
•
•
•
Patches are small
•
even if they’re on a curved surface, we can think of them as being plane
View variation for a plane patch
•
Plane patches look different in different views
Pinhole camera (F+P, p31)
Orthographic camera (F+P, p33)
Views of plane patches in perspective
cameras induce homographies
Views of plane patches in orthographic
cameras induce a special subset of
homographies
Interest points and local descriptions
•
•
Find localizable points in the image
•
e.g. corners - established technology,
eg find image windows where there tend to be strong edges going in
several different directions
•
Build at each point
•
•
•
a local, canonical coordinate frame
Euclidean+scale
Affine
Do this by searching for a coordinate frame within which some predicate
applies
E.g. Rotation frame from orientation of gradients
E.g. Rotation + scale orientation of gradients, maximum filter response
a representation of the image within that coordinate frame
this representation is invariant because frame is covariant
•
•
•
•
•
Example: Lowe, 99
•
•
Find localizable points in the image
•
find maxima, minima of response to difference of gaussians
over space
over scale
using pyramid
•
•
•
Build at each point
•
•
a Euclidean + scale coordinate frame
scale from scale of strongest response
rotation from peak of orientation histogram within window
Representation
SIFT features
•
•
•
Lowe’s SIFT features
Fig 7 from:
Distinctive image features from scale-invariant keypoints
David G. Lowe, International Journal of Computer Vision, 60, 2 (2004), pp. 91-110.
From Lowe, 99, Object Recognition from Local Scale-Invariant Features
Mikolaczyk/Schmid coordinate frames
Matching objects with point features
•
Voting
•
•
•
each point feature votes for every object that contains it
object with most votes wins
Startlingly effective (see figures)
Figure from “Local grayvalue invariants for image retrieval,” by
C. Schmid and R. Mohr, IEEE Trans. Pattern Analysis and
Machine Intelligence, 1997 c 1997, IEEE as used in Forsyth +
Ponce, p 609
Employ spatial relations
Figure from “Local grayvalue invariants for
image retrieval,” by
C. Schmid and R. Mohr, IEEE Trans. Pattern
Analysis and Machine Intelligence, 1997 c
1997, IEEE as used in Forsyth + Ponce, p 609
ASL translation
ASL
Grandma
•
•
Notice:
•
•
•
intra-class variation (timing, role of hands)
carefully dressed narrator
low resolution
Would like to spot words in text
Grandpa
Wordspotting allows translations
Wordspotting
•
Difficulties
•
•
•
•
building a discriminative word model for a large vocabulary is hard
need lots of examples of each word
building a generative word model is hard, too
no widely available pronunciation dictionary
very large number of features
next slide
•
•
•
Strategy
•
•
build spotters for some words (base words)
now use the output of those spotters as features for other words
uses less training data
because features are discriminative
•
•
Features
ASL translation
Logistic Regression
•
•
•
•
Build a parametric model of the posterior,
•
p(class|information)
•
log(P(1|data))-log(P(0|data))=linear expression in data
•
•
maximum likelihood on examples
problem is convex
•
linear
For a 2-class problem, assume that
Training
Classifier boundary
Multiclass classification
•
Many strategies
•
•
1-vs-all
for each class, construct a two class classifier comparing it to all other
classes
take the class with best output
if output is greater than some value
Multiclass
log(P(i|features))-log(P(k|features))=(linear expression)
many more parameters
harder to train with maximum likelihood
still convex
•
•
•
•
•
•
•
Forsyth Farhadi 06
Pose consistency
•
A match between an image structure and an object
structure implies a pose
•
we can vote on poses, objects
From Lowe, 99, Object Recognition from Local Scale-Invariant Features
From Lowe, 99, Object Recognition from Local Scale-Invariant Features
Kinematic grouping
•
•
Assemble a set of features to present to a classifier
•
which tests
appearance
configuration
whatever
•
•
•
Classifier could be
•
•
•
•
handwritten rules (e.g. Fleck-Forsyth-Bregler 96)
learned classifier (e.g. Ioffe-Forsyth 99)
likelihood (e.g. Felzenszwalb-Huttenlocher 00)
likelihood ratio test (e.g. Leung-Burl-Perona 95; Fergus-Perona-Zisserman
03)
Pictorial structures
•
For models with the right form, one can test “everything”
•
•
•
•
•
•
•
•
model is a set of cylindrical segments linked into a tree structure
model should be thought of as a 2D template
segments are cylinders, so no aspect issue there
3D segment kinematics implicitly encoded in 2D relations
easy to build in occlusion
putative image segments are quantized
=> dynamic programming to search all matches
What to add next? (DP deals with this)
Pruning? (Irrelevant)
Can one stop?
(Use a mixture of tree models, with missing segments marginalized
out)
Known segment colour - Felzenszwalb-Huttenlocher 00
Learned models of colour, layout, texture - Ramanan Forsyth 03, 04
•
•
•
•
•
Figure from “Efficient Matching of Pictorial Structures,”
P. Felzenszwalb and D.P. Huttenlocher, Proc. Computer Vision and Pattern Recognition
2000, c 2000, IEEE as used in Forsyth+Ponce, pp 636, 640
Constellations of parts
Fischler & Elschlager 1973
Yuille ‘91
Brunelli & Poggio ‘93
Lades, v.d. Malsburg et al. ‘93
Cootes, Lanitis, Taylor et al. ‘95
Amit & Geman ‘95, ‘99
Perona et al. ‘95, ‘96, ’98, ’00
Agarwal & Roth ‘02
Generative model for plane templates
(Constellation model)
Foreground model
Gaussian shape pdf
based on Burl, Weber et al. [ECCV ’98, ’00]
Gaussian part appearance pdf
Gaussian
relative scale pdf
log(scale)
Prob. of detection
0.8
Clutter model
Uniform shape pdf
Gaussian background
appearance pdf
0.75 0.9
Uniform
relative scale pdf
log(scale)
Poission pdf on #
detections
Figure after Fergus et al, 03; see also Fergus et al, 04
Star-shaped models
•
•
•
Features generated at parts
•
•
at image points that are conditionally independent given part location
with appearance that is conditionally independent given part type
•
•
•
very like a pictorial structure
inference is dynamic programming
localization easy
Part locations are conditioned on root
Easy to deal with
•
Other types of model
•
•
We’ve already seen a tree-structured model!
•
(pictorial structure)
•
•
because there is no conditional independence
means fewer features
Complete models are much more difficult to work with
Constellation models
•
•
Learning model
•
•
•
•
•
on data set consisting of instances, not manually segmented
choose number of features in model
run point feature detector
each response is from either one “slot” in the model, or bg
this known, easy to estimate parameters
parameters known, this is easy to estimate
missing variable problem -> EM
•
•
search for allocation of feature instances to slots that maximizes likelihood
ratio
detect with likelihood ratio test
•
•
Detecting instance
Typical models
Motorbikes
Figure after Fergus et al, 03; see also Fergus et al, 04
Spotted cats
Summary of results
Dataset
Fixed scale
experiment
Scale invariant
experiment
Motorbikes
7.5
6.7
Faces
4.6 of results
Summary
4.6
Airplanes
9.8
7.0
Cars (Rear)
15.2
9.7
Spotted cats
10.0
10.0
% equal error rate
Note: Within each series, same settings used for all datasets
Figure after Fergus et al, 03; see also Fergus et al, 04
Caution: dataset is known to have some quirky features
LOTS of BIG collections of images
Corel Image Data
40,000 images
Fine Arts Museum of San Francisco
83,000 images online
Cal-flora
20,000 images, species information
News photos with captions
(yahoo.com)
1,500 images per day available from
yahoo.com
Hulton Archive
40,000,000 images (only 230,000 online)
internet.archive.org
1,000 movies with no copyright
TV news archives
(televisionarchive.org, informedia.cs.cmu.edu)
Several terabytes already available
Google Image Crawl
>330,000,000 images (with nearby text)
Satellite images
(terrarserver.com, nasa.gov, usgs.gov)
(And associated demographic information)
Medial images
(And associated with clinical information)
* and the BBC is releasing its video archive, too;
and we collected 500,000 captioned news images;
and it’s easy to get scanned mediaeval manuscripts;
etc., etc.,
Imposing order
•
•
•
•
Iconic matching
•
•
child abuse prosecution
managing copyright (BayTSP)
•
Browsing for:
web presence for museums (Barnard et al, 01)
home picture, video collections
selling pictures
Current, practical applications
Clustering
•
•
•
Maybe applications
Searching
•
•
scanned writing (Manmatha, 02)
collections of insects
•
a face gazetteer (Miller et al, 04)
Building world knowledge
Maybe applications
Search is well studied
•
•
Metadata indexing
•
keywords, date of photo, place, etc.
•
query by example with
global features
(e.g. Flickner et al. 95, Carson et al. 99, Wang 00, various entire
conferences)
local features
(e.g. Photobook - Pentland et al 96; Blobworld - Carson et al, 98)
relevance feedback
(e.g. Cox et al 00; Santini 00; Schettini 02; etc.)
query by class
naughty pictures
(eg Forsyth et al. 96, 99; Wang et al. 98; Chan et al 99)
Content based retrieval
•
•
•
•
•
•
•
•
•
What will users pay for?
•
Work by Peter Enser and colleagues on the use of photø
movie collections
(Enser McGregor 92; Ornager 96; Armitage Enser 97; Markkula Sormunen
00; Frost et al 00; Enser 00)
Typical queries:
What is this about?
•
“… smoking of kippers…”
“The depiction of vanity in painting, the
depiction of the female figure looking in
the mirror, etc.”
“Cheetahs running on a greyhound course
in Haringey in 1932”
Annotation results in complementary words and pictures
Query on
“Rose”
Example from Berkeley
Blobworld system
Annotation results in complementary words and pictures
Query on
Example from Berkeley
Blobworld system
Annotation results in complementary words and pictures
Query on
“Rose”
and
Example from Berkeley
Blobworld system
Cluster
found
using
only text
Cluster
found
using
only blob
features
Adjacent clusters found using both text and
blob features
Input
Image
processing*
“This is a picture of the
sun setting over the sea
with waves in the
foreground”
Language
processing
sun sky waves sea
Each blob is a large
vector of features
• Region size
• Position
• Colour
• Oriented energy
(12 filters)
• Simple shape
features
* Thanks to Blobworld team [Carson, Belongie, Greenspan, Malik], N-cuts team [Shi, Tal, Malik]
FAMSF Data
83,000 images online, we clustered 8000
Searching
Compute P(document | query_items)
query_items can be words, features, or both
Natural way to express “soft queries”
Related retrieval work: Cascia, Sethi, and Sclaroff, 98; Berger and
Lafferty, 98; Papadimitriou et al., 98
Query: “river tiger” from 5,000 Corel images
(The words never occur together.)
Retrieved items: rank order P( document | query)
Search
Compute P(document | query_items)
Related retrieval work: Cascia, Sethi, and Sclaroff, 98; Berger and Lafferty, 98; Papadimitriou et al., 98
Query: “water sky cloud
Retrieved items:
”
Pictures from Words (Auto-illustration)
Text Passage (Moby Dick)
“The large importance attached to the
harpooneer’s vocation is evinced by
the fact, that originally in the old
Dutch Fishery, two centuries and
more ago, the command of a whaleship …“
Extracted Query
large importance attached fact old
dutch century more command
whale ship was person was divided
officer word means fat cutter time
made days was general vessel
whale hunting concern british title
old dutch ...
Retrieved Images
Auto-annotation
•
Predict words from pictures
•
•
•
Obstacle:
Hoffman’s model uses document specific level probabilities
Dodge
smooth these empirically
•
•
Attractions:
•
•
•
easy to score
large scale performance measures (how good is the segmenter?)
possibly simplify retrieval (Li+Wang, 03)
Keywords
GRASS TIGER CAT FOREST
Predicted Words (rank order)
tiger cat grass people water bengal
buildings ocean forest reef
Keywords
HIPPO BULL mouth walk
Predicted Words (rank order)
water hippos rhino river grass
reflection one-horned head
plain sand
Keywords
FLOWER coralberry LEAVES
PLANT
Predicted Words (rank order)
fish reef church wall people water
landscape coral sand trees
Train a system of svm classifiers, one per word but penalize that matrix for rank,
after Rennie+Srebro 05
The latent space reveals scenes because it is good at word prediction and takes
appearance into account
Loeff Farhadi Forsyth??
It was there and we didn’t
It was there and we predicted it
It wasn’t and we did
Loeff Farhadi Forsyth??
Scene
CD #
(rough proxy)
Loeff Farhadi Forsyth??
Loeff Farhadi Forsyth??
News dataset
•
Approx 5e5 news images, with
captions
•
•
•
•
Easily collected by script from Yahoo
over the last 18 months or so
Mainly people
•
•
politicians, actors, sportsplayers
long, long tails distribution
•
some images have many (resp. few)
faces, few (resp. many) names (cf.
Srihari 95)
Face pictures captured “in the
wild”
Correspondence problem
President George W. Bush makes a statement in the Rose Garden
while Secretary of Defense Donald Rumsfeld looks on, July 23,
2003. Rumsfeld said the United States would release graphic
photographs of the dead sons of Saddam Hussein to prove they
were killed by American troops. Photo by Larry Downing/Reuters
President George W. Bush makes a statement
in the Rose Garden while Secretary of Defense
Donald Rumsfeld looks on, July 23, 2003.
Rumsfeld said the United States would release
graphic photographs of the dead sons of
Saddam Hussein to prove they were killed by
American troops. Photo by Larry Downing/
Reuters
Data examples
Doctor Nikola shows a fork that was removed
from an Israeli woman who swallowed it while
trying to catch a bug that flew in to her mouth, in
Poriah Hospital northern Israel July 10, 2003.
Doctors performed emergency surgery and
removed the fork. (Reuters)
President George W. Bush waves as he leaves the
White House for a day trip to North Carolina, July
25, 2002. A White House spokesman said that
Bush would be compelled to veto Senate legislation
creating a new department of homeland security
unless changes are made. (Kevin Lamarque/
Reuters)
Process
•
•
•
•
•
•
Extract proper names
•
rather crudely, at present
Detect faces
•
Scale
with Cordelia Schmid’s face detector, (Vogelhuber Schmid
00)
44773 big face responses
Rectify faces
•
by finding eye, nose, mouth patches, affine transformation
34623 properly rectified
Kernel PCA rectified faces
Estimate linear discriminants
Now have (face vector; name_1,...., name_k)
27742 for k<=4
Rectification
•
SVM based local feature detectors, affine transformation
Kernel PCA
•
•
Find principal components - high variance features - in an
abstract feature space by computing eigenvectors of the
kernel matrix
Problems:
•
•
•
•
kernel matrix dimension (number of data items, approx 40, 000)
can’t compute all, many kernel matrix entries
incomplete cholesky no help
(column operations involve touching every face)
Nystrom approximation works, because matrix has very low rank
•
mxm
NxN
Approximate C as:
Building a face dictionary
•
•
Compute linear discriminants
•
•
Now break correspondence with modified k-means
•
•
•
•
using single name, single face data items
we now have a set of clusters
assign face to cluster with closest center,
chosen from associated names
recompute centers, iterate
using distance in LD space
•
Now recompute discriminants, recluster with modified kmeans
US President George W. Bush (L) makes
remarks while Secretary of State Colin Powell
(R) listens before signing the US Leadership
Against HIV /AIDS , Tuberculosis and Malaria
Act of 2003 at the Department of State in
Washington, DC. The five-year plan is
designed to help prevent and treat AIDS,
especially in more than a dozen African and
Caribbean nations(AFP/Luke Frazza)
German supermodel Claudia Schiffer gave birth to
a baby boy by Caesarian section January 30,
2003, her spokeswoman said. The baby is the first
child for both Schiffer, 32, and her husband, British
film producer Matthew Vaughn, who was at her
side for the birth. Schiffer is seen on the German
television show 'Bet It...?!' ('Wetten Dass...?!') in
Braunschweig, on January 26, 2002. (Alexandra
Winkler/Reuters)
British director Sam Mendes and his
partner actress Kate Winslet arrive at
the London premiere of 'The Road to
Perdition', September 18, 2002. The
films stars Tom Hanks as a Chicago hit
man who has a separate family life and
co-stars Paul Newman and Jude Law.
REUTERS/Dan Chung
and yields a useful little NLP module, too